KastorFlowUpdated October 17, 2026
AI for research: what the numbers say
How researchers use AI, and how far to trust it.
Most researchers now use AI: 58% in Elsevier's 2025 survey and 84% in Wiley's, mostly to find and summarize studies. Accuracy is the catch: GPT-4 fabricated 18% of the academic citations it wrote in a 2023 test. Grounding helps: when models only summarize a document they are given, the best stay under 4% errors. That is the case for a source-grounded notebook such as Gemini Notebook (formerly NotebookLM), which answers from the papers you add and cites them.
We make KastorFlow, a Chrome extension for Gemini Notebook. The studies on this page aren't ours; each number links to its source.
58%
of researchers use AI tools in their work, up from 37% in 2024 (3,200+ researchers)
84%
of researchers used AI in 2025, up from 57% in 2024 (2,430 researchers)
61%
use it to find and summarize new studies, the top research use; 51% for literature reviews
18%
of GPT-4's academic citations were fabricated, and 24% of its real ones had errors
1.8%
errors for the best model when summarizing a document it was given; the worst reached 24.2%
How accurate ChatGPT's academic citations were
Fabricated: share of all citations. With errors: share of the real ones. Source: Walters & Wilder, Scientific Reports 2023, 636 citations.
What that means for an AI research tool
| What the research shows | What to look for | Gemini Notebook | KastorFlow adds |
|---|---|---|---|
| GPT-4 fabricated 18% of its citations | Answers only from sources you chose, with the passage | Every answer cites your sources | An audit that scores every source in a notebook (Pro+) |
| 61% use AI to find and summarize studies | Finds sources on a question and lets you choose | Fast Research and Deep Research | Builds a whole notebook from a topic; you keep what you want (free: 5 a month) |
| Research spans many projects | Searches everything you have read | Chat works inside one notebook | One library with Ctrl+K search and tags across every notebook (free) |
| Fields change every month | Keeps sources current | No scheduled refresh | Refreshes sources and imports feeds on a schedule (Pro) |
| A free notebook holds 50 sources | Uses the room well | Up to 600 on paid Google plans | Finds duplicates and deletes in bulk (free); merges notebooks (Pro) |
How to research with AI and keep it accurate
- Start one notebook per question, from your own papers or from a topic. KastorFlow's Build research notebook gathers candidates for you to review.
- Ask questions, and open the cited passage before you rely on an answer.
- Check the notebook for weak or off-topic sources. KastorFlow's audit scores each one (1 a month free, no limit on Pro+).
- Keep the sources current with a scheduled refresh (KastorFlow Pro).
- When a new question touches old reading, search every notebook at once with Ctrl+K.
KastorFlow adds these steps to Gemini Notebook. Free to install, no account needed.
Add to ChromeSources
- Elsevier Researcher of the Future 2025
- Wiley ExplanAItions 2025
- Walters & Wilder, Scientific Reports 2023
- Vectara hallucination leaderboard, October 16, 2026
Checked October 17, 2026. Download every number with its source (CSV) · CC BY 4.0
Use these numbers
Quote any number with a link to this page. Suggested credit: KastorFlow, “AI for research: what the numbers say,” https://kastorflow.com/stats/ai-for-research-statistics, updated October 17, 2026.
Questions
What is the best AI tool for research?
For working from papers and reports you have chosen, a source-grounded notebook such as Gemini Notebook (formerly NotebookLM): it answers only from your sources and cites the passage, which matters because general chatbots fabricated 18% to 55% of academic citations in tests. KastorFlow adds topic-to-notebook research builds, search across every notebook and scheduled source refresh.
Can I trust AI-generated citations?
Not without checking. In a 2023 test, 55% of GPT-3.5's and 18% of GPT-4's academic citations were fabricated, and 24% of GPT-4's real citations had errors. Tools that answer from documents you supply and link each claim to a passage reduce the problem; open the passage before you use a claim.
How do researchers use AI?
Mostly to find and summarize new studies (61%) and for literature reviews (51%), according to Elsevier's 2025 survey of more than 3,200 researchers, in which 58% used AI tools in their work.
Is NotebookLM good for research?
Yes, for analyzing sources you have collected: up to 50 per notebook on the free plan and up to 600 on Google AI Ultra, with every answer cited. It is not a search engine for all literature; use its Fast or Deep Research to find sources, then check them.